4.5 KiB
Transport Layer and Unattended Repair
Scope: eliminate a repeated wrong-layer repair loop and publish the existing Test desktop workflows. No automated tests, console test calls, MDI interactions, browser acceptance or performance benchmark were executed. Only necessary build, publication, static metadata and process observation were used.
Why Work Repeated
Cycles 8131/8132 (Web) and 8133/8134 (Test) added or restated native registrations without new results. Actual static gaps were transport, not native. The raw and compressed scheduler prompts incorrectly told the agent that missing IDs meant missing native registration. The compressed prompt also discarded per-ID gap stages. This sent repeated repairs to the wrong files even when the diagnostic was accurate.
Both prompt forms now share REPAIR_LAYER_POLICY and implementation_blockers from scheduler_context.py. They retain per-ID stage/reason/repair and distinguish native, transport/worker, artifacts/build and mounted UI. transport does not request more native markers. Small web/src JSON dispatch tables are now tracked as source edits. No readiness guard was disabled; empty success and textual implementation_ready remain insufficient.
Test Module Delivered
The three MDI commands and console-output command already had executable native and browser-host handlers. They use TestRequest/test.inspectRegistry, not the DesktopRequest naming convention assumed by discovery. A four-entry JSON catalog now drives actual dispatch and explicit scoped discovery of this existing path. Discovery still requires native support, the matching module endpoint, actual Worker/Facade handler names and, for MDI, the mounted host orchestration.
This revealed genuine build blockers previously hidden behind missing routes: the target linked pthread dependencies without pthread/pool settings, and used fc_web_test_initialize before declaration. Both were repaired. Five workers are prestarted because the console command launches five threads before joining. Source implementations were not replaced and no diagnostic command was run.
Final receipt: test-transport-handoff-pipeline.json, four selected records ready. The independent batch did not overwrite the active scheduler manifest. Frontend and changed native artifacts were published for user use.
Web Is Still Incomplete
Update 2026-09-13: the four-command wrong-runtime transport described below has now been replaced and built. See web-document-view-routing-20260913.md and web-document-view-handoff-pipeline.json. The following paragraphs retain the original diagnosis, not instructions to repeat the repaired transport.
The Web adapter sends records without module, rejected by the stateful request guard. Adding that field alone would route them to stateful-family.worker.ts, which loads DocumentProbe's legacy stateful fixture, not WebViewCommands. The Web helper lives in a different runtime and has no shared active Part document. Therefore its catalog was deliberately not registered as a working transport.
Per-command source findings for AlignToSelection, BoxElementSelection, BoxSelection and DrawStyle identify the actual mismatch and next implementation. Fix the view host/current document/typed endpoint together; do not add a module field and call the fixture, invent a document, or report native markers as a functional handoff.
Timings
| Step | Seconds |
|---|---|
| Initial preflight rejection | 0.071 |
| Required CMake regeneration | 29.679 |
| Compile attempt exposing missing declaration | 2.876 |
| Final module registry | 0.102 |
| Final capability registry | 8.125 |
| Final preflight | 0.071 |
| Final native build | 18.077 |
| Final artifact sync | 1.018 |
| Completion metadata, not runtime acceptance | 0.936 |
| Final native pipeline total | 28.329 |
| Frontend publication | 8.905 |
The scheduler is gracefully reloaded after the current turn to load the corrected prompts. Existing histories and cooldowns are preserved; elapsed cool periods or process survival are not evidence of restored throughput. A later progress window is needed before claiming that the overall migration is healthy or complete.
Reload observed: watchdog restart count 171, new scheduler PID 2969630, cycle 8143 running, scheduler source hash matches the edited source. The current batch has moved to TestProgress1-4, not the four already delivered MDI/console commands. Those new commands are not claimed complete by this handoff. Source inventory refresh reused 496 files and reread five; no history or cooldown was erased.